A hybrid approach for bioprocess state estimation using NIR spectroscopy and a sigma-point Kalman filter

Abstract

Typically available online measurements in cultivations do not yield enough information to build up an optimization-based control for which model-based estimation techniques are necessary. More specifically, real-time information about the concentrations of substrates and biomass is the key to controlling and optimizing cultivations in a bioreactor. In addition to established techniques such as the measurement of dissolved oxygen, off-gas composition, and the amount of spent correction fluids for pH control, inline near-infrared spectroscopy (NIR) offers the possibility of gathering information online to estimate those concentrations, without the need for manual sampling. The NIR data can be transformed via the widely applied partial least squares (PLS) regression to estimate substrate and biomass concentrations. However, the spectra are corrupted by disturbances, such as gas bubbles from sparging, that can result in noisy and defective spectra and thus may lead to highly imprecise measurements. For feedback control, such imprecisions can lead to wrong decisions with respect to feeding rates, which may be hard to compensate for at a later stage. Moreover, information about operating conditions and physical constraints, such as feeding profiles and kinetic relations between substrates that could be used to rectify erroneous predictions, are disregarded in the PLS model. In this paper, we therefore present a hybrid approach of NIR spectroscopy and nonlinear, model-based state estimation to enable improved quality in the online estimation of substrates and biomass in aerated yeast cultivations. The feeding rates, off-gas concentrations, NIR spectra, and knowledge about the dynamics of the process are integrated seamlessly in a sigma-point Kalman filter (SPKF). The latter is formulated such that it is able to handle physical boundaries, which is especially important for fed-batch cultivations. In order to evaluate the hybrid approach, it is compared to different scenarios typically applied in fermentations. One uses only NIR data with a PLS regression model. The other works with an SPKF, but without NIR information. It is shown that the presented hybrid method outperforms the methods applied separately for the estimation of biomass, glucose, ethanol, ammonium, and phosphate in cultivations of a new, genetically modified strain of Saccharomyces cerevisiae. As only a few cultivations are needed for calibration, a reliable online state estimation is available in the early stage of development of this process.